Of all the methods for geometric rectification of remote sensing imagery, the Collinearity Equation Model

is usually considered to have the best accuracy. Nevertheless, when the Collinearity Equation Model based on GCPs

(ground control points) is used to compute the elements of inner and exterior orientation, the coefficient matrix

condition of the normal equation often becomes deteriorative, which greatly affects the accuracy of the orientation

elements. In this paper, a new method for geometric rectification based on neural network is proposed. Experiments

show that, under the precondition that a certain number of GCPs serve as the training data, the neural network of BP

and RBF can perform well in geometric rectification of remote sensing imagery and reach higher accuracy than the

Collinearity Equation Model. Besides, the neural network can eliminate the influence of GCPs with gross error, and

hence can better improve the efficiency.